> ## Documentation Index
> Fetch the complete documentation index at: https://critiqor-71f5274a.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Runtime Observation: How Critiqor Watches Your Agent

> Critiqor records tool calls, retries, failures, and cost signals during your agent's run — giving you factual behavioral evidence beyond the final output.

Runtime observation is the foundation of how Critiqor evaluates your AI agent. Instead of judging your agent only by its final output, Critiqor watches what your agent actually *does* during a run — recording tool calls, retries, failures, timing, and cost signals as they happen. This gives you a factual record of agent behavior, not just a snapshot of the result it reported.

This matters because final outputs can look correct even when the path to get there was inefficient, error-prone, or unreliable. Runtime observation surfaces the difference.

## What gets captured

Critiqor records the following during an observation session:

* **Tool calls and their outcomes** — every tool your agent invokes, whether the call succeeded, failed, or returned unexpected results.
* **Retry attempts and loops** — repeated attempts at the same action, including cases where your agent retries without apparent progress.
* **Failures and error events** — exceptions, timeouts, and other failure signals that occurred during the run, even if the agent ultimately recovered.
* **Execution timing and sequencing** — the order in which actions were taken and how long each phase of the run took.
* **Cost signals** — redundant tool calls, unnecessary repeated work, and other patterns that inflate execution cost without adding value.
* **Agent behavior patterns across the session** — how your agent's behavior evolved over the course of a run, including any escalation in retries or error frequency.

## Runtime vs final-answer evaluation

|                               | Runtime observation                                                | Final-answer evaluation                                                   |
| ----------------------------- | ------------------------------------------------------------------ | ------------------------------------------------------------------------- |
| **What it shows**             | What the agent actually did during execution                       | What the agent reported as its output                                     |
| **Catches silent failures**   | Yes — failures and retries are recorded even if the agent recovers | No — a recovered failure looks identical to a clean run                   |
| **Reveals inefficiency**      | Yes — redundant calls and loops are visible in the evidence        | No — output quality says nothing about how much unnecessary work was done |
| **Reflects real reliability** | Yes — based on observed behavior across the full session           | Partial — a correct answer can come from an unreliable process            |

Runtime evidence reveals what the agent actually did. Final-answer evaluation only shows what the agent reported.

## When observation is active

Observation starts when you run `critiqor monitor openclaw` and ends when you run `critiqor finalize`. During this window, Critiqor runs passively in the background, recording runtime evidence from your agent's session. When you finalize, Critiqor uses that evidence to produce the reliability report for the run.

<Info>
  Critiqor is passive — it does not modify your agent's behavior or intercept its responses. Your agent runs exactly as it would without Critiqor present.
</Info>

For a step-by-step walkthrough of the observation workflow, see [Observation Workflow](/guides/observation-workflow).